基于迁移学习的交通场景车辆实时检测算法

Real-time Vehicle Detection Algorithm for Traffic Based on Transfer Learning

为实现精确实时的车辆检测,本文算法基于迁移学习思想,以深度学习实时检测算法YOLOv2为基础。使用在大规模数据集上预训练得到的分类模型初始化YOLOv2卷积神经网络,搜集交通场景车辆图片并标注后输入该网络利用反向传播进行微调,从而得到最终的车辆检测模型。测试结果表明,本文算法在包含300张车辆图片的测试集中MAP达到0.788,每帧检测平均耗时15ms,满足工程应用实时性要求。

In order to achieve accurate and real-time vehicle detection, our algorithm bases on transfer learning and YOLOv2（a deep learning real-time detection algorithm）. The convolution neural network of YOLOv2 is initialized by the classification model which is pre-trained on largescale data sets. The collected vehicle images of traffic scene are marked then input to the network, using backpropagation to fine-tune the network to obtain the final vehicle detection model. The test results show that our algorithm achieves MAP of 0.788 in the test set containing 300 vehicle images.For each frame test, average cost time is 15 ms, meeting the real-time requirements of engineering applications.